90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-081
Media · Australia · 2024 · Undisclosed AI image manipulation

9News Melbourne aired an altered image of MP Georgie Purcell, reportedly exposing her midriff and enlarging her chest, and attributed the change to Photoshop's automation

By Ellie Harris · Filed Image broadcast 29 January 2024

Alleged: Nine Entertainment (Nine Network; 9News Melbourne) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

9News Melbourne aired an altered image of MP Georgie Purcell, reportedly exposing her midriff and enlarging her chest, and attributed the change to Photoshop's automation

What happened

It was reported that on 29 January 2024 a 9News Melbourne bulletin, covering the Victorian government’s position on duck hunting, used a still image of Georgie Purcell, a member of the Victorian Legislative Council for the Animal Justice Party and a prominent campaigner against the practice. In the broadcast image, the dress Purcell had worn in the original photograph appeared as separates that exposed her midriff, and her chest appeared enlarged. Purcell posted the two images side by side, said the network had edited her body and her outfit, and described the result as sexist, adding that she could not imagine it happening to a male colleague. She later spoke on ABC RN Breakfast about AI-augmented imagery and the way it falls on women.

Nine apologised. The director of news in Melbourne, Hugh Nailon, said the graphics department had sourced an image of Purcell online and resized it to fit the bulletin’s specifications, and that during that process automation in Adobe Photoshop had produced a picture that was not consistent with the original, which did not meet the network’s editorial standards. The tool at issue was Photoshop’s generative fill, and its generative expand option, which extends a picture beyond its original frame by inventing plausible content to fill the new space, powered by Adobe’s Firefly model. Adobe, whose software had been named as the cause, responded that any changes of this kind to the image would have required human intervention and approval, that its generative features do not run and publish alterations on their own. The two accounts did not reconcile: Nine described an automated step that overreached, and the tool’s maker described a process a person has to direct and accept. What was never produced publicly was the record that would have settled it: the source image, the edits applied to it, and the person who approved it before it went to air.

What an auditable version would have shown

The dispute between Nine and Adobe came down to a missing record. A newsroom that alters a photograph holds, or should hold, the evidence of what it did: the original file, each tool applied to it, the parameters of any generative step, and the sign-off of the editor who released it. With that record, the claim that automation alone changed the image is either shown to be true or shown not to be, in seconds, from the file’s own history rather than from competing statements days later. This is not hypothetical tooling. Adobe’s own Content Authenticity Initiative, begun in 2019, attaches content credentials to an image, a tamper-evident log of what device made it and what edits were performed, precisely so that an altered picture arrives carrying its own provenance. That system already existed, and whatever was used to make this image, the picture that went to air did not carry those credentials.

Where the gap was

An image of a real person was altered and published, and no record was produced to show what had been changed or by whom. A ConstraintGate encodes the standing editorial rule as a check that runs before publication: a generative alteration of a real person’s photograph is not released without disclosure and a named human approval, so the rule is enforced at the moment of use rather than recalled in an apology. A ConductRecord preserves the picture’s history, the source file, the tools and generative parameters applied, and the editor who signed it off, as a signed and reviewable entry, so that responsibility for what a tool and a person did together can be read from the file instead of contested between the outlet and the software vendor.

What governance should have looked like

When a tool and a person both handle an image, “the automation did it” can only be assessed against a record, and here there was none to check. Two decisions sit behind the published picture: running the generative tool, and releasing what it produced. Best practice would be for a media organisation that publishes an altered image of a real person to be able to show, from its own history, what was done to the image and who approved its release, so that responsibility rests with a named person rather than an unresolved dispute with the software vendor.

Failure Pattern: a generative tool altered a real person’s published image, and no provenance record could show what was changed, by what, or who approved it.

Governance Principle: a published image of a real person should carry a verifiable record of its source and every edit applied, so an alteration is disclosed rather than deniable.

The reference implementation of ConstraintGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Nine Entertainment (Nine Network; 9News Melbourne) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →